Zombie Stats: AI Numbers Often Lack Real Sources

Trakkr Research said it followed 1,285 AI-cited statistics back to where they came from, and found that fewer than half of those trails reached a study that supported the number. The work matters because AI assistants are now stating statistics as fact, sometimes with confident citations that lead to blogs or dead ends rather than research. The study was published on the Hacker News front page and is dated October 29, 2026.

What did the Trakkr Research study find?

The report described a source-checking project. According to the source at https://trakkr.ai/trakkr-research/zombie-stats, researchers asked ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews 200 questions that called for a number. They then clicked the citations and kept clicking. In one example, all five assistants answered a question about how many salespeople give up after one follow-up. Each gave 44%. Only ChatGPT said the evidence behind it was unclear.

The team followed 1,285 cited statistics and traced each trail, sometimes up to five pages deep. The reported outcomes were stark. Forty percent reached a study that said the same thing. Eight percent reached a study, but the number or its context had changed. Forty percent stopped at a source the team could not check, because it was locked, gone or silent on the number. Eleven percent led nowhere. The last 1% had pages that would not load.

The sample was drawn from statistics each assistant cited. ChatGPT had 108 citations, Perplexity 277, Gemini 300, Claude 300 and Google AI Overviews 300. The five-assistant rate weighted each assistant by how many cited statistics it quoted, so it described the statistics people were shown. Where a page named a source without a link, the team searched for it.

The report said ChatGPT did best, quoting the fewest numbers and sourcing the most. ChatGPT quoted 1.3 statistics per answer, against 6.8 for Claude. The other assistants quoted more, with rates that the researchers said were too close to rank. The overall finding was that many AI numbers may arrive without a solid trail.

Why do AI citations often point to blogs instead of studies?

A central finding was that AI often cited the copy, not the study. The researchers said the page an AI cited had the figure in 82% of cases. Only 21% of the pages cited for a number were the research itself. More than half were company blogs. Add news stories and statistics roundups, and most citations landed on pages built to collect figures from elsewhere, often without a link back.

The starting point of the trail mattered. Research firms were 13% of first citations and 62% of those were sourced. Government sources were 7% and 61% sourced. Company blogs were 54% and 37% sourced. News was 10% and 37% sourced. Stats roundups were 8% and 35% sourced. Even research firms fell short 38% of the time, often because the full report sat behind a form.

The report suggested that the trouble was often which page an AI cited, not whether the AI invented the number outright. Most numbers were not necessarily zombies at the first stop. The problem was the path after the citation.

How do numbers change as AI answers travel?

The study also examined what happened after a trail reached a study. Of the statistics that got there, 17% had changed on the way. The digits moved only about one time in four. The context changed often: what was measured, the year, the source, or who was asked.

The changes logged were: number changed 26%, what was measured changed 20%, year changed 17%, source changed 16%, who was asked changed 14%, and place changed 4%. One example involved a 2004 study by Sillence and colleagues. It watched 15 women judge health websites and found 94% of comments about rejecting or mistrusting sites were about design. Over time, the finding was repeated as 94% of first impressions being design-related. Google AI Overviews in 2026 stated it as fact.

The original study was narrow. The AI answer was broad. That gap was one form of drift the research flagged.

What does this mean for AI answers and research?

The findings suggest that AI assistants may be useful for generating language around numbers, but not always reliable for grounding those numbers. The researchers followed trails and found that source quality, context and citation quality varied widely.

For creators, developers and researchers, the practical lesson is not to stop using AI statistics. It is to verify them. A number that arrives quickly may be convenient, memorable and shared. It may also be unmoored from its original study.

The study noted that AI models judged trails and that human checks corrected some conclusions. It also said ChatGPT cited fewer numbers per answer than some rivals, while still showing mixed sourcing performance. The broader lesson is that citation does not automatically mean provenance. The report called the problem a matter of trust, not only accuracy.

How can readers and creators check AI stats before sharing?

Readers can ask an assistant for the primary source, then follow the link. They can ask for the study, sample size, date and exact wording. Creators can compare multiple AI outputs and treat repeated numbers with suspicion. Researchers can look for raw answers and methodology. A useful check is to ask the assistant for the original study rather than the summary.

For people working in AI creation spaces, this is relevant to Grafy. A creator who wants to make content about AI, or a developer who wants to test claims, can use Grafy Chat to put a question to multiple models side by side and ask each one for its sources. A researcher can use Grafy Reader to work through papers and primary documents while checking claims.

The Trakkr Research piece does not show that Grafy has a specific tool for tracing AI citations. It does show why the problem matters to anyone who uses AI for research, content or product claims.

Frequently asked questions

What is a zombie stat?

A zombie stat is a number that AI assistants repeat confidently even when the citation trail leads to no study or a weak source. The Trakkr Research study followed AI answers and found many numbers stopped at blogs, roundups or uncitable pages.

How many AI statistics survived source checks?

According to the study, about 40% of traced AI statistics reached a study that said the same thing. Another 8% reached a study but the number or context had changed. The rest could not be fully checked or led nowhere.

What should creators do with AI stats?

They should ask for the primary source, check the date and context, and trace the citation before repeating it. AI can help generate hypotheses, but a statistic used in content should be verified against the original study. That habit protects trust and accuracy.

---

Source: Hacker News — front page — Zombie stats: Where the numbers AI quotes come from